DeerFlow

ByteDance's DeerFlow started as a deep research framework and, with the ground-up 2.0 rewrite released in June 2026, became a general super-agent harness that orchestrates sub-agents, memory, and sandboxes through extensible skills. Built on LangGraph and LangChain, it spawns sub-agents for parallel execution, persists memory through DeerMem, mem0, or OpenViking backends, and executes code in local, Docker, or Kubernetes sandboxes, with MCP server integration for external tools. The bundled web UI covers chat, thread and memory management, and a file workspace, while skills extend to image and video generation, slide deck creation, agentic browser control of interactive websites, and messaging channels including Telegram, Slack, Feishu, and WeChat. Deployment uses Docker or a local setup with the uv package manager for the Python backend and Node.js 22 with pnpm for the frontend, plus an interactive make setup wizard, and you bring your own LLM provider keys. Everything is MIT licensed, and at about 78.2k GitHub stars it ranks among the most complete open agent stacks to come out of the deep research wave.